Measured execution costs
Powered by a vLLM-based layerwise profiler, with reproducible TTFT, TPOT, and latency comparisons against recorded Llama and Qwen workloads.
Heterogeneous HW
Mix GPU, CPU, CXL, and PIM tiers. Drop in any hardware target via the per-hardware CSV bundle format.
Disaggregated serving
First-class TP / PP / EP / DP+EP support across multiple instances, with wave-synchronized expert collectives on multidimensional ASTRA-Sim topologies.
Real workloads
Drive the simulator with ShareGPT-style traces or closed-loop agentic sessions (sub-requests + tool calls) for SWE-bench-style scenarios.
Recognition
Three publications. Three awards.
3
Publications
2
Best Paper Awards
1
Distinguished Artifact
- IISWC 2024Best Paper AwardDistinguished Artifact Award
- ISPASS 2026Best Paper Award